Department of Computer & Information Science

Departmental Papers (CIS)

University of Pennsylvania Year 

VC-Dimension of Exterior Visibility

Volkan Isler∗ Sampath K. Kannan†
Kostas Daniilidis‡ Pavel Valtr∗∗

∗ University of Pennsylvania
† Universityof Pennsylvania, [email protected]
‡ University of Pennsylvania, [email protected]
∗∗ Charles University

Copyright 2004 IEEE. Reprinted from IEEE Transactions on Pattern Analysis and
Machine Intelligence, Volume 26, Issue 5, May 2004, pages 667-671.
Publisher URL: <a href=”http://ieeexplore.ieee.org/xpl/tocresult.jsp?isNumber=28505&page=1”
>http://ieeexplore.ieee.org/xpl/tocresult.jsp?isNumber=28505&page=1</a>

This material is posted here with permission of the IEEE. Such permission of the IEEE does
not in any way imply IEEE endorsement of any of the University of Pennsylvania’s products
or services. Internal or personal use of this material is permitted. However, permission to
reprint/republish this material for advertising or promotional purposes or for creating new
collective works for resale or redistribution must be obtained from the IEEE by writing to
[email protected]. By choosing to view this document, you agree to all provisions of
the copyright laws protecting it.

This paper is posted at ScholarlyCommons.

http://repository.upenn.edu/cis papers/18
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 26, NO. 5, MAY 2004 667

VC-Dimension of Exterior Visibility of image based rendering, we might have a prior rough map of the
environment but we need to obtain a detailed reconstruction with
a range sensor. In other cases, we have a geometric model, but we
Volkan Isler, Student Member, IEEE, do not know anything about the color or texture of an object. In all
Sampath Kannan, these cases, it is important that the rendered environment does not
Kostas Daniilidis, Member, IEEE, and have any holes because of originally uncovered areas.
Pavel Valtr We are not going to address here the equally important
problem of unknown environments or objects related to model
building tasks. Several algorithms exist for exploring unknown
Abstract—In this paper, we study the Vapnik-Chervonenkis (VC)-dimension of set
systems arising in 2D polygonal and 3D polyhedral configurations where a subset
environments [8], building upon fundamental results in the on-line
consists of all points visible from one camera. In the past, it has been shown that the traversal of graphs [9], or on Markov processes for modeling
VC-dimension of planar visibility systems is bounded by 23 if the cameras are partially known dynamic scenes [10]. Significant contributions
allowed to be anywhere inside a polygon without holes [1]. Here, we consider the have been also achieved in the Next Best View planning problem
case of exterior visibility, where the cameras lie on a constrained area outside the [11], [12], [13] for surface acquisition. However, the results proven
polygon and have to observe the entire boundary. We present results for the cases in this paper have implications for the unknown case as well:
of cameras lying on a circle containing a polygon (VC-dimension= 2) or lying outside
Choosing sensor locations randomly is a method frequently used
the convex hull of a polygon (VC-dimension= 5). The main result of this paper
concerns the 3D case: We prove that the VC-dimension is unbounded if the
for sensor placement in unknown environments [14]. The
cameras lie on a sphere containing the polyhedron, hence the term exterior visibility. VC-dimension theory enables us to answer the question: How
many random samples (sensors) do we need in order to cover a given
Index Terms—VC-dimension, sensor placement, sampling, visibility. region? Our results, together with the theory of
-nets, provide an
upper bound to the answer to this question when omni-directional
æ cameras are used as sensors.
It is well known [2] that the minimal guard coverage problem is
1 INTRODUCTION
NP-hard. To study the existence of approximation algorithms, we
IMAGINE a known 3D polyhedral environment where a set of can consider minimal guard coverage as an instance of the set-cover
cameras has to be placed in such a way that every point in the problem. The general version of minimum set-cover cannot be
environment is visible. The 2D version is known as the art gallery approximated with a ratio better than log n. However, we do not
problem [2], [3], [4] and sufficiency results exist for several know whether any set-cover instance can be realized as a visibility
versions of this problem. For example, bn3 c cameras can cover any system. A powerful interface between set-cover and the particular
simple polygon. However, such results are inapplicable in robotic geometric setup is the Vapnik-Chervonenkis (VC) dimension, which
and image-based rendering applications where the environments enables us to quantitatively bound how general a set system is.
can be very complex with millions of vertices. A further In this paper, we present new bounds on the VC-dimension for
application is placing antennas for line-of-sight broadband com- three instances of the problem. In Section 2, we formalize the
munication [5]. Imagine that backbones end at each neighborhood problem statement and describe our contribution to the state of the
and that communication inside 1km can be achieved with line-of-
art. Next, we deal with 2D configurations and prove new bounds.
sight laser beams that can carry from 10Mbps up to 1.25 Gbps
Specifically, we prove that, if the cameras are restricted to lie on a
bandwidth. Assuming that a consumer can put a receiving antenna
circle, the VC-dimension is 2 (Section 3.1) and, if they remain
at the window of her studio or even on a kiosk in a street, the
outside the convex hull of the object, the VC-dimension is 5
coverage problem becomes a visibility problem where the cameras
(Section 3.2). In Section 4, we prove the main result that the VC-
become arrays of distributing antennas.
dimension for 3D configurations is unbounded. We conclude with
In this paper, we consider aspects of the problem of minimizing
the number of viewpoints without sacrificing the goal of complete a summary in Section 5.
visibility. The particular scenarios we are addressing are surveil-
lance, object inspection, and image based rendering. In the case of 2 PROBLEM STATEMENT
surveillance, we need a complete coverage at any time so that no
event will be missed. This is the reason why coverage with one 2.1 VC-Dimension and Set-Cover
mobile guard (shortest watchman route—solvable in polynomial A set system is a pair ðX; RÞ where X is a set and R is a collection
time [6]) is not applicable. In case of object inspection [7], we know of subsets R
X.
the prior geometry of an object, and we need the minimal number Definition 1. Given a set system ðX; RÞ, let A be a subset of X. We say
of views so that the object will be checked regarding defects. In this A is shattered by R if 8Y
A; 9R 2 R such that R \ A ¼ Y . The
scenario, the object might be placed on a turntable and we ask then VC-dimension of ðX; RÞ is the cardinality of the largest set that can
for the minimal number of rotations. The objects might even be be shattered by R [15].
medical organs which have to be imaged from very few view-
points of an endoscope guided by a robot manipulator. In the case The VC-dimension, introduced first in supervised learning for
pattern classification, plays an important role also in randomized
and geometric algorithms [16], [17]. As an example, we mention
. V. Isler, S. Kannan, and K. Daniilidis are with the Department of
Computer and Information Science, University of Pennsylvania, Levine the related notion of an
-net without going into definitions: If the
Hall, 3330 Walnut Street, Philadelphia, PA 19104-6389. set system ðX; RÞ has a constant VC-dimension d, then with high
E-mail: {isleri, kannan, kostas}@cis.upenn.edu. probability, a small number ðOðd
log 1
ÞÞ of points sampled from X
. P. Valtr is with the Institute for Theoretical Computer Science and intersects all the subsets in R whose sizes are greater than
jXj
Department of Applied Mathematics, Charles University, Malostranské
(such a sample is called an
-net). Another useful property is that,
nám. 25, 118 00 Praha 1, Czech Republic.
E-mail: [email protected]. if ðX; RÞ has a constant VC-dimension d, then the number of
subsets in R is Oðnd Þ where n ¼ jXj. A related result [18] that
Manuscript received 3 Oct. 2002; revised 26 May 2003; accepted 3 Oct. 2003.
Recommended for acceptance by R. Sharma. implicitly deals with
-nets is the existence of polyhedra that
For information on obtaining reprints of this article, please send e-mail to: require arbitrary number of guards even if every point inside the
[email protected], and reference IEEECS Log Number 117501. polyhedron can see an area equal to
fraction of the total interior
0162-8828/04/$20.00 ß 2004 IEEE Published by the IEEE Computer Society
668 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 26, NO. 5, MAY 2004

area, using the notion of
-good polygons introduced in path

planning [19].
Given a set system, the minimum set-cover problem asks for a
S
minimum cardinality set S
R such that R2S R ¼ X. The hitting
set problem is the dual of set-cover and its decision version reads:
We are given a set X and a collection of sets R where each R 2 R is
a subset of X. We are also given a number k. The question is
whether there is a subset H  X such that jHj  k and for each
R 2 R, R \ H 6¼ ;.
Both problems are known to be NP-complete and can be
Fig. 1. Interior visibility extends to exterior visibility by turning the polygon inside out.
approximated to within a log factor of the maximum set sizes (in
either the primal or the dual system) and not better [20]. For sets
systems with finite VC-dimension d, however, Brönnimann and approximated better than a factor of log n. However, as mentioned
Goodrich presented an algorithm which returns a solution whose in the first section, it is not clear that general set-cover instances
size is at most Oðd  log OP T  OP T Þ [21]. Here, OP T is the can be realized by visibility systems.
cardinality of the optimal solution. This is a significant improve- Valtr proved that the VC-dimension of the system
ment on the previous approximation, when the cardinality of the ðP ; fV ðxÞ j 8x 2 P gÞ, where P is a simple polygon and V ðxÞ
optimal solution is smaller than the cardinality of the maximum set is the visibility polygon of point x in P , is bounded between 6
size in the set system. and 23 [1]. This result applies to all 2D configurations we
consider, simply by turning the polygon inside-out (see Fig. 1,
2.2 Visibility and Set-Cover also [3]). He also established a bound of Oðlog hÞ for polygons
In this paper, we will address the problem of minimal guard with h holes. A similar result for visibility under angle and
coverage or camera placement. An instance of the camera distance constraints has been obtained in [23].
placement problem is: Given a polygon or a polyhedron P and a On the other hand, it is not clear how to exploit a bounded
specification of possible camera locations, find a minimum set- VC-dimension to obtain an improved approximation algorithm.
cover of the system ðP ; fV ðci ÞgÞ, where V ðci Þ is the set of points Approximation algorithms for minimum guard coverage have
visible on P from camera ci and the index i varies over all possible been considered [22], [24], [23] for different versions of the
camera locations. The definition of V ðci Þ can capture any optical problem. However, there is still a gap between the inapproxim-
constraints on what a camera can see. We will refer to the ability results and existing algorithms.
specification of possible camera locations as a setup. We say a set S As mentioned before, the minimum set-cover of set systems
S
of cameras cover P if ci 2S V ðci Þ ¼ P . Depending on the applica- with bounded VC-dimension can be approximated within a
tion, P may refer to the boundary of the object or it may be logarithmic factor of the optimal value [21]. However, this by
extended to include the interior of the environment as well. itself does not improve on the existing log n approximations, as the
Camera placement can also be seen as a particular case of the optimum can be as big as n=3 [2]. Nevertheless, this algorithm was
hitting set problem. The set X is the set of possible camera used in [23] to get rid of the dependency of the approximation
locations. For each point p on the boundary of the polyhedron P , factor to the number of samples (rather than the number of
there is a set Rp consisting of all camera locations that can see p. vertices). In fact, obtaining a constant approximation algorithm for
The hitting set problem assumes a finite set X and we have to guard placement in polygons without holes is one of main open
implicitly deal with this issue when we attempt to pose camera problems in the field of art galleries.
placement as such a problem. Typically, we do so by using a For the setup where cameras are restricted to lie on a circle, an
sampling technique or by showing that a finite set of extremal approximation algorithm that returns at most one more camera
points need to be considered. than the minimum necessary is presented in [25]. A similar
Throughout this paper, we will represent cameras with their algorithm can be found in [26] for the polygon separation problem.
projection centers ci and say that ci sees the point p 2 P if the only Placing cameras outside the convex hull of a polygon is related to
intersection of the line segment pci with P is p. We extend the hitting lines with points [27], [28]. Unfortunately, there are no
notion of visibility to sets as follows: We say that a camera sees a improved algorithms for this restricted case either.
set of points ! if it can see all the points in !. The following
notation will be useful for VC-dimension proofs: Let Pm ¼
fp1 ; . . . ; pm g be m points on P . We say that camera c sees the
3 RESULTS ON PLANAR CONFIGURATIONS
subset !
Pm if c can see all points in ! but no point in Pm n !. 3.1 2DCIRCLE
By the VC-dimension of a setup, we will refer to the Consider a setup in which we want to cover a polygon P using
VC-dimension of the maximum number of points that can be cameras restricted to a circle C around P . Note that not all
shattered over all instances of the camera placement problem for a polygons are completely visible from a circle outside. In this
specific setup. For example, if there are no restrictions on cameras section and the next, we restrict ourselves to the points on P that
and we want to cover simple polygons, we would like to find the are visible from at least one point on the circle. Polygons that are
VC-dimension of the set system ðP ; fV ðci ÞgÞ as P varies over the entirely visible from the viewing circle are called externally visible
set of all simple polygons. Therefore, in order to give a lower polygons [29].
bound m on the VC-dimension of a setup, it suffices to present one Definition 2. We define 2DCIRCLE as a setup where a set of cameras
instance where m points are shattered, but, for an upper bound m, whose locations are restricted to a circle C are to cover a simple
one needs to show that there exists no instance such that m þ 1 polygon that is contained in C.
points can be shattered.

2.3 Related Work on VC-Dimension We need the following technical lemma before proving our
main proposition. Its proof is omitted because it is straightforward.
In general, it is possible to consider visibility systems as set
systems and camera placement as a set-covering problem [22]. The Lemma 3. Each point p on the polygon P is visible along a continuous
general version of the minimum set-cover problem cannot be arc on the circle C and nowhere else.
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 26, NO. 5, MAY 2004 669

Fig. 3. Case 1 of Lemma 6: The fpi g-cell gets two marks and the ;-cell gets one
mark. The unbounded fpi g-cell is marked .
Fig. 2. Points fp1 ; p2 g can be shattered by four cameras. Each camera is labeled
with the subset it can see. In this figure, the polygon P is 4ABC.
Thus, we may suppose that each point pi lies inside convP .
Whenever an i-ray is disjoint from a j-ray, i 6¼ j, for technical
We now prove the following proposition. reasons we “create” a new, “abstract” cell and associate it with
Proposition 4. The VC-dimension of 2DCIRCLE is exactly 2. this pair of disjoint rays. The total number of “real” and
“abstract” cells is exactly 71, provided no three rays intersect in
Proof. For any m points on the polygon P , the m visibility arcs have
a single point (otherwise, it is smaller).
2m endpoints and, hence, there are only 2m combinatorially
distinct camera positions. Since 2m cameras are necessary to We suppose that every camera c! is placed in an unbounded
cell whenever it is possible. We remove the camera c; and 63
shatter m points, we need 2m  2m which is only true for m  2.
cameras in 63 different cells remain. We say that a cell is empty,
The lower bound is proven by the example in Fig. 2, where
if it contains no camera. All “abstract” cells are empty.
the points p1 and p2 are shattered by the four cameras shown.t u
We now describe a procedure which distributes 18 auxiliary
3.2 2DCONVEX marks 1; 1; 1; 2; 2; 2; . . . ; 6; 6; 6 in some of the empty cells in
Let us now relax the restriction on camera locations so that we such a way that at most two marks are placed in one cell. It will
allow cameras anywhere outside the convex hull of the polygon. follow that at least 18=2 ¼ 9 cells are empty.
For each i ¼ 1; . . . ; 6, the three marks i are distributed as
Definition 5. We define 2DCONVEX as a setup where a set of cameras
follows. We say that an i-ray and a j-ray form an i-pair, if they
located outside the convex hull of a simple polygon P are to cover P .
are disjoint. We distinguish several cases.
The upper bound on the VC-dimension of 2DCONVEX slightly
Case 1: There is at most one i-pair. One of the i-rays is
increases, but it is still a small constant significantly less than the
intersected by all j-rays for j 6¼ i. Hence, its endpoint sees no
upper bound for the general case, 23. point pj ; j 6¼ i. We place two marks i in the adjacent fpi g-cell
Lemma 6. The VC-dimension of 2DCONVEX is less than or equal to 5. and one mark i in the adjacent ;-cell (see Fig. 3). Note that the
cell with two marks i is empty because it is bounded and there
Proof. Suppose that Q ¼ fp1 ; . . . ; p6 g is a set of six points shattered
is an unbounded fpi g-cell in this case.
on a polygon P . Let C be the boundary of convP , the convex
hull of P , and let E be the exterior of C, i.e., E ¼ R2 n ðconvP Þ. Case 2: There are two i-pairs. One mark i is placed in each of
the two abstract cells associated with the i-pairs. The remaining,
For each i ¼ 1; . . . ; 6, a point of E sees the point pi if and only if
third mark i is put in an ;-cell chosen as follows.
it lies between a certain pair of disjoint open half-lines
emanating from C. We call these two open half-lines the i-rays Subcase 2a: No point pj ; j 6¼ i, is visible from the endpoint e of one
of the i-rays. In this case, we put the third mark i in the ;-cell
and refer to the point on C from which they emanate as their
adjacent to e.
endpoints. The 12 i-rays, i ¼ 1; . . . ; 6, partition E into cells.
Formally, cells are maximal connected components of the plane Subcase 2b: Condition in Subcase 2a is not valid. In this case,
note that each i-ray must be responsible for exactly one i-pair.
after the removal of convP and of the 12 rays.
Otherwise, one of the i-rays would intersect all other j-rays and
We can also obtain the cells in the following way. One by
the endpoint of this i-ray would only see pi and this is covered
one and in an arbitrary order, we remove the 12 rays from E.
in Subcase 2a. Let i1 and i2 be the two i-rays. Suppose that the
After the removal of the first ray there is still one component
endpoint of i1 sees some pj ; j 6¼ i. Then, the wedge of visibility
(cell). Each other ray divides every intersected cell into two
of pj must intersect i1 in a bounded interval starting at the
smaller cells. Thus, if the ray intersects k of the previously
endpoint. For all other k 6¼ j, the k-rays intersect i1 and it is
removed rays, its removal increases the
number of cells at most immediate that this intersection is a bounded interval again.
by k þ 1. Since there are at most 12 2  6 ¼ 60 intersections
Thus, as we go to infinity along i1 , we must have the case that
between the 12 rays (the two i-rays are disjoint for each i), the
only pi is visible on one side and no point is visible on the other
number of cells in the final arrangement is at most 60 þ 11 ¼ 71.
side. We put the third mark i to this unbounded ;-cell adjacent
It follows from the construction that cameras placed at
to i1 . Note that this ;-cell is not adjacent to C, as can be verified
different locations in the same cell see the same subset ! of
from the analogous reasoning for i2 .
Q ¼ fp1 ; . . . ; p6 g. We then call the cell an !-cell. To get a
contradiction, it suffices to show that there are no 26  1 ¼ 63 Case 3: There are three or more i-pairs. We put a mark i in three
cells seeing distinct nonempty subsets of Q. abstract cells associated with distinct i-pairs.
First, suppose that one of the points pi lies on C. Then, the We now check that no cell receives more than two marks.
two i-rays are parts of lines tangent to C. It is easily verified that Any abstract cell is associated with a pair of disjoint rays and
each j-ray, j 6¼ i, is disjoint from at least one of the two i-rays in may receive at most two marks corresponding to these two
this case. Whenever a j-ray is disjoint from an i-ray, j 6¼ i, this rays. A fpi g-cell receives at most two marks i. An ;-cell adjacent
decreases the number of cells in the final arrangement by 1. to C may receive at most two marks corresponding to the two
Thus, the number of cells is at most 71  10 ¼ 61 in this case, rays having the endpoints on the boundary of the cell. An ;-cell
which is not enough. nonadjacent to C may receive at most two marks corresponding
670 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 26, NO. 5, MAY 2004

Fig. 4. The small picture on the left shows a polygon with five shattered points fp1 ; p2 ; p3 ; p4 ; p5 g. The figure on the right is a detail showing the intersection pattern of the
i-rays. Each label 2; 3; 4; 5 indicates the cardinality of the subset seen from the region.

to rays forming unbounded edges of the cell. All other cells catalogue of events (structural changes in the image of the
receive no marks. polyhedron), they construct the viewing space partition, VSP,
Hence, no cell receives more than two marks. Therefore, which is a partition of the viewpoint space into maximal regions of
there are at least 18=2 ¼ 9 cells with at least one mark. They are constant aspect. They also present tight bounds for the number of
all empty. Thus, at most 71  9 ¼ 62 cells are nonempty. They regions in VSP. They show that the size of the VSP for a general
cannot contain all the 63 cameras c! ; ;  !
Q. u
t (i.e., nonconvex) polyhedron under orthographic projection is
Note that relaxing the camera locations from 2DCIRCLE to
ðn6 Þ and their model for the orthographic projection is exactly the
2DCONVEX indeed increases the VC-dimension, as we see in the same as 3DSPHERE with S at infinity.
following lemma.
Lemma 11. The VC-dimension of 3DSPHERE is Oðlog nÞ.
Lemma 7. The VC-dimension of 2DCONVEX is greater than or equal to 5.
Proof. Let Pm ¼ fp1 ; . . . ; pm g be any m points to be shattered on a
Proof. An instance where five points are shattered is presented polyhedron. If we define an aspect as appearance/disappear-
in Fig. 4. u
t ance of pi ; i ¼ 1; . . . ; m, and restrict the camera locations to a
The result of this section is summarized in the following sphere that contains the polyhedron, we can use the catalogue
theorem, which follows from Lemma 6 and Lemma 7: of events in [30] to show that the size of the VSP for this new
Theorem 8. The VC-dimension of 2DCONVEX is exactly 5. notion of aspects is still
ðn6 Þ. However, in order to shatter m
points, one needs 2m distinct partitions. Since we must have
Remark. If we further remove the restriction that the cameras are 2m ¼ Oðn6 Þ, we have m ¼ Oðlog nÞ, which gives us the desired
outside the convex hull, then the best known bound is 23 and upper bound. u
t
the reader is referred to [1].
4.2 Lower Bound
4 RESULTS ON 3D CONFIGURATIONS In this section, we show that the upper bound log n on the
VC-dimension of 3DSPHERE is indeed tight, our main result is
In this section, we consider the following setup which arises in stated in the following lemma for Theorem 10.
typical tele-immersive applications:
Lemma 12. For any given m, there exists a polyhedron P with
ð2m Þ
Definition 9. We define 3DSP HERE as a setup where we are given vertices such that there are m marked points on P that can be
a polyhedron P, and a viewing sphere S such that P is totally shattered from 2m cameras on the viewing sphere S.
contained in S.
Proof. We take a regular simplex (tetrahedron) T inside a viewing
We show that even when the viewing region is restricted to a
sphere S. Let F be a facet (2-dimensional face) of T . We take a
sphere, there are polyhedra with n vertices such that
ðlog nÞ
set M of m points slightly above the center of F . Further, we
points on the polyhedron can be shattered from the viewing sphere
place 2m cameras c! ; !
M, on S visible from each point of M.
S that contains P. Namely, we prove the following theorem:
We choose them so that no camera lies on a line determined by
Theorem 10. The VC-dimension of 3DSPHERE is
ðlog nÞ. two points of M.
In the next two sections, we present the upper and lower Then, for any p 2 !
M, we choose a segment sðp; !Þ
bounds for the VC-dimension of 3DSPHERE, in Lemmas 11 and 12, having one endpoint on the face F and intersecting the segment
respectively. pc! and no other segments p0 c!0 ; p0 2 M; !0
M. Clearly, we
may choose the segments sðp; !Þ pairwise disjoint. Then, we
4.1 Upper Bound replace each sðp; !Þ by a very thin simplex Sðp; !Þ growing out
In this section, we present an upper bound on the VC-dimension of of the face F and intersecting the segment pc! and no other
3DSPHERE. In [30], Platinga and Dyer define aspects as changes in segments p0 c!0 ; p0 2 M; !0
M. If the simplices Sðp; !Þ are thin
the topology of the image of a polyhedron. After presenting a enough, then they are pairwise disjoint.
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 26, NO. 5, MAY 2004 671

Let P be the union of T with the m  2m1 simplices Sðp; !Þ. [12] K. Kutulakos and C. Dyer, “Recovering Shape by Purposive Viewpoint
Adjustment,” Int’l J. Computer Vision, vol. 12, pp. 113-136, 1994.
Then, P shatters the set M and has m  2mþ1 þ 4 vertices. It has a [13] V. Hlavac, A. Leonardis, and T. Werner, “Automatic Selection of Reference
facet F 0 obtained from F by the removal of m  2m1 triangular Views for Image-Based Scene Representations,” Proc. European Conf.
holes. We arbitrarily triangulate F 0 . If we now perturb the Computer Vision, pp. 526-535, 1996.
[14] N. Bulusu, J. Heidemann, and D. Estrin, “Adaptive Beacon Placement,”
vertices of F 0 , then P may be changed to a simplicial Proc. Int’l Conf. Distributed Computing Systems, 2001.
polyhedron P such that F 0 is replaced by triangular facets [15] V.N. Vapnik and A.Y. Chervonenkis, “On the Uniform Convergence of
corresponding to the chosen triangulation of F 0 . u
t Relative Frequencies of Events to Their Probabilities,” Theory of Probability
and Its Applications, vol. 16, no. 2, pp. 264-280, 1971.
Proof of Theorem 10. Theorem 10 is a direct consequence of [16] J. Matousek, “Geometric Computation and the Art of Sampling,” IEEE
Lemmas 11 and 12. t
u Symp. Foundations of Computer Science—Tutorial, 1998.
[17] P.K. Agarwal and S. Sen, “Randomized Algorithms for Geometric
Optimization Problems,” Handbook of Randomization, P. Pardalos et al.,
eds., Kluwer Academic Publishers, 2003.
5 CONCLUSION [18] P. Valtr, “On Galleries with No Bad Points,” Discrete & Computational
Geometry, vol. 21, pp. 193-200, 1999.
Visibility of a polygonal or polyhedral configuration can be cast as [19] L.E. Kavraki, J.-C. Latombe, and R. Motwani, “Randomized Query
a set system with subsets defined by the visibility region of each Processing in Robot Path Planning,” J. Computer and System Sciences,
vol. 57, pp. 50-60, Aug. 1998.
camera. The minimal guard coverage problem can then be
[20] P. Slavik, “A Tight Analysis of the Greedy Algorithm for Set Cover,”
formulated as a minimum set-cover problem. The constraints J. Algorithms, vol. 25, pp. 237-254, 1997.
imposed by the geometry of the setup can be captured with the [21] H. Brönnimann and M. Goodrich, “Almost Optimal Set Covers in Finite
VC-Dimension,” Discrete & Computational Geometry, vol. 14, no. 4, pp. 463-
VC-dimension of the visibility set system. It was known [1] that the 479, 1995.
upper bound of the VC-dimension for polygons is 23. In this paper, [22] S. Ghosh, “Approximation Algorithms for Art Gallery Problems,” Proc.
we improved this bound for two cases of exterior visibility: Canadian Information Processing Soc. Congress, 1987.
[23] H. Gonzalez-Banos and J. Latombe, “A Randomized Art-Gallery Algorithm
cameras on a circle containing the polygon and cameras outside for Sensor Placement,” Proc. ACM Symp. Computational Geometry (SoCG ’01),
the convex hull of a polygon. The circle case has significant pp. 232-240, 2001.
[24] S. Eidenbenz, C. Stamm, and P. Widmayer, “Inapproximability Results for
practical implications because it minimizes the number of stations
Guarding Polygons and Terrains,” Algorithmica, vol. 31, pp. 79-113, 2001.
of a turn-table or a laser-scanner pedestal in 3D-modeling and [25] V. Isler, S. Kannan, and K. Daniilidis, “VC-Dimension of Exterior Visibility
object inspection. The placement of cameras outside the convex of Polyhedra,” Technical Report: MS-CIS-01-34, Computer and Information
Science, Univ. of Pennsylvania, 2001.
hull is significant in surveillance and image-based rendering. For [26] J.S.B. Mitchell and S. Suri, “Separation and Approximation of Polyhedral
this case, as well as the case of arbitrary placement inside a Surfaces,” Proc. Third ACM-SIAM Symp. Discrete Algorithms (SODA),
polygon, the existence of approximation algorithms with constant pp. 296-306, 1992.
[27] V.S.A. Kumar, S. Arya, and H. Ramesh, “Hardness of Set Cover with
ratio is still an open problem. Intersection 1,” Automata, Languages and Programming, pp. 624-635, 2000.
In the 3D case, we showed that for any n, a polyhedron with [28] B. Brodén, M. Hammar, and B.J. Nilsson, “Guarding Lines and 2-Link
n vertices can be constructed such that the VC-dimension of its Polygons is Apx-Hard,” Proc. 13th Canadian Conf. Computational Geometry,
pp. 45-48, 2001.
exterior visibility is
ðlog nÞ. [29] C. Hernando, M.E. Houle, and F. Hurtado, “On Local Transformation of
Polygons with Visibility Properties,” Theoretical Computer Science, vol. 289,
pp. 919-937, Oct. 2002.
ACKNOWLEDGMENTS [30] W. Plantinga and C. Dyer, “Visibility, Occlusion, and the Aspect Graph,”
Int’l J. Computer Vision, vol. 5, no. 2, pp. 137-160, 1990.
The authors highly appreciate the comments by the anonymous
reviewers, which substantially improved the readability of the
paper. The first three authors are grateful for support through the . For more information on this or any other computing topic, please visit our
following grants: US National Science Foundation NSF-IIS- Digital Library at www.computer.org/publications/dlib.
0083209, NSF-IIS-0121293, NSF-CCR-98-20885, NSF-CCR-01-
05337, MURI DAAH-19-02-1-03-83, and a Penn Research Founda-
tion grant. Research by P. Valtr was supported by project
LN00A056 of The Ministry of Education of the Czech Republic.

REFERENCES
[1] P. Valtr, “Guarding Galleries Where No Point Sees a Small Area,” Israel
J. Math., vol. 104, pp. 1-16, 1998.
[2] J. O’Rourke, Art Gallery Theorems And Algorithms. Oxford Univ. Press, 1987.
[3] T. Shermer, “Recent Results in Art Galleries,” Proc. IEEE, vol. 80, pp. 1384-
1399, 1992.
[4] J. O’Rourke, “Visibility,” Handbook of Discrete and Computational Geometry.
J.E. Goodman and J. O’Rourke eds., ch. 25, pp. 467-480, CRC Press LLC,
1997.
[5] A. Acampora, “Last Mile by Laser,” Scientific Am., vol. 287, pp. 48-53, July
2002.
[6] W. Chin and S. Ntafos, “Shortest Watchman Routes in Simple Polygons,”
Discrete and Computational Geometry, vol. 6, no. 1, pp. 9-31, 1991.
[7] R. Sharma and J. Molineros, “Computer Vision-Based Augmented Reality
for Guiding Manual Assembly,” PRESENCE: Teleoperators and Virtual
Environments, vol. 6, pp. 292-317, 1997.
[8] F. Hoffmann, C. Icking, R. Klein, and K. Kriegel, “The Polygon Exploration
Problem,” SIAM J. Computing, vol. 31, no. 2, pp. 577-600, 2002.
[9] X. Deng, T. Kameda, and C. Papadimitriou, “How to Learn an Unknown
Environment i: The Rectilinear Case,” J. ACM, vol. 45, pp. 215-245, 1998.
[10] S.M. LaValle and R. Sharma, “On Motion Planning in Changing, Partially-
Predictable Environments,” Int’l J. Robotics Research, vol. 16, pp. 775-805,
1997.
[11] K. Tarabanis, P. Allen, and R. Tsai, “A Survey of Sensor Planning in
Computer Vision,” IEEE Trans. Robotics and Automation, vol. 11, pp. 86-104,
1995.